Activation Channel Scales and Calibration

mit-han-lab / smoothquant

[ICML 2023] SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models

https://arxiv.org/abs/2211.10438

MIT License

1.26k stars 150 forks source link

Open 520zw opened 1 year ago

520zw commented 1 year ago

Hi, here at generate act scales.py，‘You can download the validation dataset of the Pile at https://mystic.the-eye.eu/public/AI/pile/val.jsonl.zst’，"wget https://mystic.the-eye.eu/public/AI/pile/val.jsonl.zst" but always display "failed: Connection timed out."

llx-08 commented 1 year ago

You can find it in Huggingface: Link，but after i downloaded the dataset and use it, it output that dataset is corrupt and unusable